The verdict in three sentences
A distributor with 8,000 SKUs that drives purchasing with moving averages in its ERP ties up too much stock in slow movers and runs out of seasonal items. A custom machine learning demand forecasting model costs 50,000 to 120,000 EUR in Amsterdam in 2026 with 3 years of history (30,000 to 70,000 EUR excl. VAT for a comparable project in France), while a SaaS solution costs 1,000 to 4,000 EUR a month. The realistic goal is average stock down 10 to 20% and service level up 2 to 5 points, which quickly adds up to hundreds of thousands of euros of working capital.
Why classic methods hit their limit
ERP moving averages and seasonal coefficients treat each SKU in isolation. They ignore promotions, weather (a real factor in the Netherlands for garden, DIY or beverages), public holidays such as King's Day, product launches and substitution effects between items. Machine learning learns these effects from sales history and external variables.
| Method | Typical forecast error (WMAPE) | Handles promotions | Indicative 2026 cost |
|---|---|---|---|
| ERP moving average | 35 to 50% | No | Included in ERP |
| Seasonal exponential smoothing | 30 to 40% | Manually | Included or module 8,000 to 20,000 EUR |
| ERP vendor APS module | 25 to 35% | Partly | 30,000 to 90,000 EUR + licence |
| Machine learning forecasting SaaS | 20 to 30% | Yes | 1,000 to 4,000 EUR a month |
| Custom model (gradient boosting, time series) | 18 to 28% | Yes, with your own variables | 50,000 to 120,000 EUR |
These error rates are 2026 orders of magnitude and vary widely with catalogue type: fast movers forecast well, the long tail much less so.
Detailed budget for a custom project
Cost splits between data preparation, modelling and above all integration into replenishment proposals, without which the model is just another spreadsheet.
| Item | Budget | Duration |
|---|---|---|
| Data audit and cleaning (3 years of sales, stockouts, promotions) | 10,000 to 20,000 EUR | 3 to 4 weeks |
| Modelling and backtesting by product family | 18,000 to 40,000 EUR | 4 to 8 weeks |
| Safety stock and reorder point calculation | 6,000 to 12,000 EUR | 2 weeks |
| ERP integration (purchase proposals, Exact, AFAS, Microsoft Dynamics) | 10,000 to 35,000 EUR | 3 to 6 weeks |
| Buyer dashboard | 6,000 to 13,000 EUR | 2 weeks |
| Hosting and monthly retraining | 500 to 1,500 EUR a month | Ongoing |
| Total project | 50,000 to 120,000 EUR | 3 to 5 months |
SaaS or custom
SaaS starts faster (6 to 10 weeks) and costs 12,000 to 48,000 EUR a year. Custom becomes more attractive after 3 years of use, when you have specific business variables (customer contracts, delivery routes, retailer promotions) or want to own the model. Dutch companies can explore the WBSO R&D tax credit and the MIT innovation scheme, to be checked case by case.
Mini case study
Pieter, supply chain director of a technical supplies distributor near Amsterdam, manages 8,000 SKUs, an average stock of 6 million EUR and a 92% service level.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
- Custom project: 90,000 EUR + 12,000 EUR a year for hosting and retraining.
- Conservative assumption: average stock -12%, so 720,000 EUR of working capital released.
- Inventory carrying cost (financing, storage, obsolescence) estimated at 20% a year: annual saving of 144,000 EUR.
- Service level raised to 95%: 3 points of lost sales recovered on 40 million EUR of revenue at a 25% margin, a large extra margin potential even counting only a third of the effect.
- Payback on stock savings alone: (90,000 + 12,000) / 144,000, so about 8.5 months.
FAQ
How much history is needed?
Ideally 3 years of daily or weekly sales to capture two to three seasons. With 18 months a model works, but the gain on seasonal items is often halved.
What about past stockouts in the history?
Zero sales during a stockout is not zero demand. Cleaning reconstructs true demand, a step that often represents 20 to 30% of the budget.
Do buyers stay in control?
Yes. The model proposes, the buyer approves or adjusts. A good dashboard shows forecast versus actual so the team trusts the tool within 2 to 3 months.
What about new products?
Without history, the model relies on similar products. Error stays high (40% and above) for the first 8 to 12 weeks.
Do we need an in-house data scientist?
Not necessarily. With a maintenance contract of 500 to 1,500 EUR a month, the provider monitors model drift and retrains it monthly.
Let's scope your project. We assess your data, the SKU scope to cover and an indicative budget between SaaS and custom. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
